Torch is a scientific computing framework with wide support for machine learning algorithms that puts GPUs first. It is a wrapper written in the Lua language around the THNN library. Torch is easy to use and efficient, thanks to an easy and fast scripting language, LuaJIT, and an underlying C/CUDA implementation.
This version of Torch should not be confused with PyTorch, which is a Python wrapper around the THNN library.
PyTorch provides similar functionality but is more actively maintained. We recommend most people use PyTorch instead of this version of Torch.
See also the documentation for PyTorch on ShARC.
A GPU-enabled worker node must be requested in order to use the GPU version of this software. See Using GPUs on ShARC for more information.
Torch is available on ShARC as both Singularity images and as a module.
This software and documentation is maintained by the RSES group and GPUComputing@Sheffield. For feature requests or if you encounter any problems, please raise an issue on the GPU Computing repository.
Singularity images are self-contained virtual machines similar to Docker. For more information on Singularity and how to use the images, see Singularity.
A symlinked file is provided that always point to the latest image:
To get a bash terminal in to an image for example, use the command:
singularity exec --nv /usr/local/packages/singularity/images/torch/gpu.img /bin/bash
exec command can also be used to call any command/script inside the image e.g.
singularity exec --nv /usr/local/packages/singularity/images/torch/gpu.img th yourscript.lua
--nv flag enables the use of GPUs within the image and can be removed if the software you’re using does not use the GPU.
You may get a warning similar to
groups: cannot find name for group ID ..., this can be ignored and will not have an affect on running the image.
/shared are automatically mounted to your ShARC filestore directories. For GPU-enabled images the
/nvbin is mounted to the correct Nvidia driver version for the node that you’re using.
To submit jobs that uses a Singularity image, see Submitting Batch Jobs That Uses Singularity Images for more detail.
Paths to the actual images and definition files are provided below for downloading and building of custom images.
The following is an instruction on how to use the Torch module.
Load the Torch module which also loads anaconda 3.4, CUDA 8.0, cuDNN 5.1 and GCC 4.9.4.
module load apps/torch/nvidia-7/gcc-4.9.4-cuda-8.0-cudnn-5.1-conda-3.4
On the DGX-1, load the NCCL library optimised for the hardware:
On any other node use a generic NCCL build:
moudle load libs/nccl/generic/gcc-4.9.4-cuda-8.0
(Optional if you plan to interface with python) Create a conda environment to load relevant modules on your local user account and activate it
conda create -n torch python=3.5 source activate torch
In the interactive session or your batch script, load the relevant modules and (optionally) activate your conda environment
module load apps/torch/nvidia-7/gcc-4.9.4-cuda-8.0-cudnn-5.1-conda-3.4 #Optional source activate torch